eduzhai > Applied Sciences > Engineering >

Contrastive learning of global and local features for medical image segmentation with limited annotations

  • Save

... pages left unread,continue reading

Document pages: 18 pages

Abstract: A key requirement for the success of supervised deep learning is a largelabeled dataset - a condition that is difficult to meet in medical imageanalysis. Self-supervised learning (SSL) can help in this regard by providing astrategy to pre-train a neural network with unlabeled data, followed byfine-tuning for a downstream task with limited annotations. Contrastivelearning, a particular variant of SSL, is a powerful technique for learningimage-level representations. In this work, we propose strategies for extendingthe contrastive learning framework for segmentation of volumetric medicalimages in the semi-supervised setting with limited annotations, by leveragingdomain-specific and problem-specific cues. Specifically, we propose (1) novelcontrasting strategies that leverage structural similarity across volumetricmedical images (domain-specific cue) and (2) a local version of the contrastiveloss to learn distinctive representations of local regions that are useful forper-pixel segmentation (problem-specific cue). We carry out an extensiveevaluation on three Magnetic Resonance Imaging (MRI) datasets. In the limitedannotation setting, the proposed method yields substantial improvementscompared to other self-supervision and semi-supervised learning techniques.When combined with a simple data augmentation technique, the proposed methodreaches within 8 of benchmark performance using only two labeled MRI volumesfor training, corresponding to only 4 (for ACDC) of the training data used totrain the benchmark. The code is made public atthis https URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×